Cross-domain Ensemble Distillation for Domain Generalization

Domain generalization is the task of learning models that generalize to unseen target domains. We propose a simple yet effective method for domain generalization, named cross-domain ensemble distillation (XDED), learns domain-invariant features while encouraging model converge flat minima, which recently turned out be sufficient condition generalization. To this end, our generates an output logits from training data with same label but different domains and then penalizes each mismatch ensemble. Also, we present de-stylization technique standardizes encourage produce style-consistent predictions even in arbitrary domain. Our greatly improves capability public benchmarks image classification, cross-dataset person re-ID, semantic segmentation. Moreover, show learned by are robust against adversarial attacks corruptions.